摘要
在输电线路巡检中,关键电力部件经常会出现锈蚀、脱落等缺陷,从而造成误检、漏检问题,导致人力成本大大增加。文中针对此类问题,提出一种基于YOLOv6的改进算法,主要在neck部分增加了RepFPN结构,额外的增加了一个自底向上的路径聚合网络,从而增强输出特征的表达能力,提升模型性能,然而增加了计算成本。为了降低计算量,增加了RepVGG结构,它有着类似残差的结构,拥有丰富的梯度信息,经实验数据表明,当使用3个RepVGG结构替代C3层时,会提升1~2点精度;同时将ConvBNAct的算子进行融合后,只需要采用一个Conv+inplace activation就可以完成原有的3次Op计算,绝大多数情况下还能得到一致的数值结果,缓解了上述改进增加计算量的问题。实验表明,改进后的模型mAP_0.5:0.95提高了12%,精确度提高了3.9%,各项loss值显著降低,同时很好地检测出了电力部件各种缺陷以及输电线路上的异物。
In the patrol of transmission lines,key power components often have such defects as corrosion and shedding,resulting in false and missed detection as well as great increase of labor costs.In view of such problem,a kind of improved algorithm based on YOLOv6 is proposed in this paper.The RepFPN structure is mainly added in the neck part,and a bottom-up path aggregation network is added to enhance the expression ability of the output features and improve the performance of the model.However,it increases the computational cost.In order to reduce the amount of calculation,the RepVGG structure with a similar residual structure and rich gradient information is added.The experimental data show that when three RepVGG structures are used in replace of C3 layer,the accuracy of 1-2 points will be improved.At the same time,after the fusion of operator of ConvBNAct,only one Conv+inplace activation is needed to complete the original three times of Op calculations.In most cases,the identical numerical results can be obtained,which alleviates the problem of increasing the amount of calculation in the above improvements.The experiment shows that the improved model mAP_0.5:0.95 is increased by 12%,the accuracy is increased by 3.9%,and each loss value is significantly reduced and,at the same time,various defects of power components and foreign objects on transmission lines are well detected.
作者
游越
伊力哈木·亚尔买买提
吐松江·卡日
YOU Yue;YILIHAMU Yaermaimaiti;TUSONGJIANG Kari(School of Electrical Engineering,Xinjiang University,Urumqi 830049,China)
出处
《高压电器》
CAS
CSCD
北大核心
2024年第5期194-205,213,共13页
High Voltage Apparatus
